Information processing method, information processing device and program

Abstract

To provide an information processing method, an information processing device and a program capable of reducing cost of time, troubles, expenses, or the like for acquiring an optimal condition.SOLUTION: An information processing method is an information processing method for creating a combinatorial library of a search space obtained by making the number of parameters constituting an experimental condition the number of dimensions, generates a prediction value of a characteristic distribution in the search space, calculates a distance field from one or more experienced experimental conditions in the search space by considering the prediction value, determines the next experimental condition on the basis of the calculated distance field, acquires a measurement result obtained by measuring a sample created by using the next experimental condition that is determined, and repeats update of the prediction value one or more times on the basis of the acquired measurement result.SELECTED DRAWING: Figure 14

Description

[Technical Field]

[0001] The present invention relates to an information processing method, an information processing device, and a program. [Background technology]

[0002] Combinatorial methods have traditionally been known as a method for searching for optimal materials, compounds, or products for a particular application (hereinafter referred to as optimal solutions). For example, in the field of chemistry, combinatorial chemistry exists, in which a wide variety of compound groups (libraries) are efficiently synthesized by combining multiple raw materials based on combinatorial theory, and the resulting compound libraries are then utilized for a variety of purposes. The method of synthesizing a wide variety of compound groups at once based on combinatorial theory is called combinatorial synthesis, and is positioned as a core technology in combinatorial chemistry.

[0003] In recent years, research has been conducted into the use of AI (Artificial Intelligence) technology to search for materials that are optimal for a particular application.

[0004] For example, Patent Document 1 shown below discloses a technique for generating compounds that have high binding affinity with a target protein by training a conditional variational auto-encoder (CVAE) using a set of pairs of known proteins and compounds that are known to have high binding affinity, and inputting the target protein (distributed representation) and latent variables (sampled with random noise) into the decoder of the trained conditional variational auto-encoder.

[0005] Furthermore, Patent Document 2 discloses a technology for generating compounds having activity against a selected target using a generative model trained using a training set including compounds and biological or chemical information related to the compounds.

[0006] Furthermore, Patent Document 3 discloses a technology for identifying a desired dye material that satisfies all of multiple physical properties, using a VAE encoder trained to receive dye material information expressed in a predetermined notation as input and output latent variables in a latent space corresponding to the dye material information, a VAE decoder trained to receive any latent variable in the latent space 5 as input and output dye material information expressed in a predetermined notation, and data related to multiple physical properties of the dye material. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Patent Publication No. 2021-68410 [Patent Document 2] Patent Publication No. 2021-121927 [Patent Document 3] Patent No. 7180806 [Non-patent literature]

[0008] [Non-Patent Document 1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin (2017), Attention Is All You Need, arXiv:1706.03762 [cs.CL]. [Non-patent document 2] Rishi Bommasani et. al (2021), On theOpportunities and Risks of Foundation Models, arXiv:2108.07258 [cs.LG]. [Non-patent document 3] Tom B. Brown, et. al (2020), LanguageModels are Few-Shot Learners, arXiv:2005.14165 [cs.CL]. Summary of the Invention [Problem to be solved by the invention]

[0009] However, general combinatorial methods require the preparation and measurement of samples of all or many possible combinations of candidates, which poses the problem of the considerable time, effort (e.g., number of trials) and costs involved in creating a combinatorial library. In addition, combinations other than those of materials prepared as samples must be searched for using methods such as Bayesian optimization, which are commonly used in fields such as MI (Material Informatics). However, in order to perform sufficiently reliable searches (predictions) using these conventional search methods, a large number of measurement results are required, which presents the problem of difficulty in sufficiently reducing costs such as time, effort (e.g., number of trials) and costs.

[0010] Furthermore, the method of searching for optimal materials using AI technology, such as the one exemplified above as a conventional technology, requires a huge amount of training data to train the generative model, and it takes a long time for the generative model to learn, which poses the problem of requiring a huge amount of time, effort, and cost to obtain the optimal solution.

[0011] Therefore, an object of the present invention is to provide an information processing method, an information processing device, and a program that make it possible to reduce the cost required to obtain an optimal solution. [Means for solving the problem]

[0012] An information processing method according to one embodiment of the present invention is an information processing method for creating a combinatorial library of a search space whose dimensionality is the number of parameters that make up the experimental conditions, which generates a predicted value of a property distribution in the search space, calculates a distance field from one or more experimental conditions that have already been tested in the search space taking into account the predicted value, determines next experimental conditions based on the calculated distance field, obtains measurement results by measuring a sample prepared using the determined next experimental conditions, and repeats the process of updating the predicted value based on the obtained measurement results one or more times.

[0013] An information processing device according to one embodiment of the present invention is an information processing device connected via a predetermined network to an experimental system that prepares a sample to be measured and a measurement system that measures the sample, and generates a predicted value of a characteristic distribution in a search space whose dimensionality is the number of parameters that constitute the experimental conditions for preparing the sample, calculates a distance field from one or more experimental conditions that have already been tested in the search space taking into account the predicted value, determines next experimental conditions based on the calculated distance field, transmits the determined next experimental conditions to the experimental system, acquires measurement results obtained by measuring a sample prepared in the experimental system using the next experimental conditions in the measurement system, and updates the predicted value based on the acquired measurement results, repeating this process one or more times to create a combinatorial library of the search space.

[0014] A program according to one embodiment of the present invention is a program for causing a processor provided in an information processing device connected via a predetermined network to an experimental system for preparing a sample to be measured and a measurement system for measuring the sample, and causes the processor to perform the following operations: generate a predicted value of a characteristic distribution in a search space whose dimensionality is the number of parameters constituting the experimental conditions for preparing the sample; calculate a distance field from one or more experimental conditions that have already been tested in the search space, taking into account the predicted value; determine next experimental conditions based on the calculated distance field; transmit the determined next experimental conditions to the experimental system; acquire measurement results obtained by measuring a sample prepared in the experimental system using the next experimental conditions in the measurement system; and update the predicted value based on the acquired measurement results. [Effects of the Invention]

[0015] According to an information processing method according to an embodiment of the present invention, it is possible to reduce the number of experiments in areas of the search space where an optimal solution is likely to exist, while focusing on areas where an optimal solution is likely to exist, thereby reducing the cost required to create a combinatorial library, and thereby reducing the cost required to obtain an optimal solution.

[0016] According to an information processing device according to an embodiment of the present invention, it is possible to reduce the number of experiments in areas of the search space where an optimal solution is likely to exist, while focusing on areas where an optimal solution is likely to exist, thereby reducing the cost required to create a combinatorial library, and thereby reducing the cost required to obtain an optimal solution.

[0017] According to a program according to one embodiment of the present invention, it is possible to reduce the cost required to create a combinatorial library by conducting experiments intensively in regions in the search space where an optimal solution is likely to exist, while reducing the number of experiments in regions where an optimal solution is unlikely to exist, thereby reducing the cost required to obtain an optimal solution. [Brief explanation of the drawings]

[0018] [Figure 1] 1 is a schematic diagram showing an example of the general configuration of an MI search system as an information processing system according to an embodiment of the present invention. [Figure 2] 1 is a block diagram showing a schematic configuration example of an experimental system according to an embodiment of the present invention. [Figure 3] 1 is a block diagram showing a schematic configuration example of a measurement system according to an embodiment of the present invention. [Figure 4] 1 is a block diagram illustrating a schematic configuration example of a prediction system according to an embodiment of the present invention. [Figure 5] 10 is a flowchart illustrating an example of an experiment point determination operation when a prediction library is not used according to an embodiment of the present invention. [Figure 6] FIG. 10 is a diagram illustrating an example of a proposal distribution according to an embodiment of the present invention. [Figure 7] FIG. 6 is a diagram showing an example of candidate points selected in step S102 of FIG. 5. [Figure 8] FIG. 6 is a diagram showing an example of candidate points selected by repeatedly executing step S106 in FIG. 5 (part 1). [Figure 9] FIG. 6 is a diagram showing an example of candidate points selected by repeatedly executing step S106 in FIG. 5 (part 2). [Figure 10] FIG. 6 is a diagram showing an example of candidate points selected by repeatedly executing step S106 in FIG. 5 (part 3). [Figure 11] FIG. 6 is a diagram showing an example of candidate points selected by repeatedly executing step S106 in FIG. 5 (part 4). [Figure 12]FIG. 6 is a diagram showing an example of candidate points selected by repeatedly executing step S106 in FIG. 5 (part 5). [Figure 13] FIG. 2 is a diagram illustrating an example of a prediction library according to an embodiment of the present invention. [Figure 14] 10 is a flowchart showing an example of an experiment point determination operation when using a prediction library according to an embodiment of the present invention. [Figure 15] FIG. 15 is a diagram showing an example of candidate points selected in step S102 of FIG. 14. [Figure 16] FIG. 15 is a diagram showing an example of candidate points selected in step S206 of FIG. 14 (part 1). [Figure 17] FIG. 15 is a diagram showing an example of candidate points selected in step S206 of FIG. 14 (part 2). [Figure 18] FIG. 15 is a diagram showing an example of candidate points selected in step S206 of FIG. 14 (part 3). [Figure 19] FIG. 15 is a diagram showing an example of candidate points selected in step S206 of FIG. 14 (part 4). [Figure 20] FIG. 15 is a diagram showing an example of candidate points selected in step S206 of FIG. 14 (part 5). [Figure 21] FIG. 1 is a diagram illustrating a schematic configuration example of a first generative model according to an embodiment of the present invention. [Figure 22] FIG. 2 is a diagram illustrating the learning of a first generative model according to one embodiment of the present invention. [Figure 23] 1 is a graph showing the scaling law of the Transformer. [Figure 24] FIG. 10 is a diagram illustrating an example of the structure of a Transformer. [Figure 25] FIG. 2 is a diagram illustrating a schematic configuration example of a second generative model according to one embodiment of the present invention. [Figure 26] FIG. 10 is a diagram for explaining transfer learning when creating a second generative model according to one embodiment of the present invention. [Figure 27] FIG. 10 is a diagram illustrating a schematic configuration example of a third generative model according to one embodiment of the present invention. [Figure 28]10 is a graph for explaining the learning status of a third generative model according to one embodiment of the present invention. [Figure 29] FIG. 10 is a diagram illustrating a schematic configuration example of a fourth generative model according to one embodiment of the present invention. [Figure 30] 10 is a graph showing prediction results of a third generative model after training according to an embodiment of the present invention. [Figure 31] FIG. 1 is a diagram for explaining an overview of the cross-validation method. [Figure 32] FIG. 10 is a diagram for explaining the results of evaluating the third generative model according to one embodiment of the present invention using a cross-validation method. [Figure 33] FIG. 10 shows an example of a prediction library obtained by combinatorial mapping of non-experimental locations using a third generative model according to one embodiment of the present invention (temperature combinatorial no. 1). [Figure 34] FIG. 10 shows an example of a prediction library obtained by combinatorial mapping of non-experimental locations using a third generative model according to one embodiment of the present invention (temperature combinatorial no. 2). [Figure 35] FIG. 10 shows an example of a prediction library obtained by combinatorial mapping of non-experimental locations using a third generative model according to one embodiment of the present invention (pressure combinatorial part 1). [Figure 36] FIG. 10 shows an example of a prediction library obtained by combinatorial mapping of non-experimental locations using a third generative model according to one embodiment of the present invention (pressure combinatorial part 2). [Figure 37] FIG. 10 is a diagram showing an example of a prediction library obtained by continuously changing the temperature and pressure input into a third generative model according to one embodiment of the present invention. [Figure 38] 1 is a flowchart illustrating an example of the operation of a generative model according to an embodiment of the present invention. [Figure 39] FIG. 2 is a diagram showing an example of a sample substrate according to an embodiment of the present invention. [Figure 40]FIG. 10 is a diagram illustrating an example of an input data arrangement when hysteresis is taken into consideration according to an embodiment of the present invention. [Figure 41] FIG. 10 is a diagram illustrating an example of an input data arrangement when hysteresis is not taken into consideration according to an embodiment of the present invention. [Figure 42] FIG. 1 illustrates an example of a combinatorial library obtained with a generative model according to an embodiment of the present invention. [Figure 43] 1 is a hardware configuration diagram illustrating an example of an information processing apparatus according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0019] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. In the following embodiments, the same components are designated by the same reference numerals, and redundant description will be omitted.

[0020] <0. Introduction> For example, in the development of devices with light-emitting layers, such as semiconductor lasers and LEDs (Light Emitting Diodes), the crystal structure of light-emitting layers formed using a combination of various materials is usually analyzed by X-ray diffraction (XRD (X-ray diffraction)). Therefore, it is easy to obtain a combination of experimental conditions during film formation (e.g., a combination of one or more parameters from the layer structure, film formation method for each layer, film formation temperature, film formation time, composition, film thickness, chamber pressure, plasma excitation frequency, dose amount, etc.) and measurement data (e.g., X-ray diffraction data) obtained by measuring the light-emitting layer formed under these experimental conditions.

[0021] When combinatorial techniques are used to search for combinations of luminescent materials, conventionally, one or more substrates (e.g., single-crystal substrates such as semiconductor wafers, hereinafter also referred to as samples) whose film thickness, composition, etc. vary depending on the position are fabricated, and X-ray diffraction patterns are measured for each position on the sample to identify the dependency of the X-ray peak on the film thickness, composition, etc. of the luminescent material. Then, the combination of materials that is considered to be optimal is determined based on the identified dependency.

[0022] However, as described above, such conventional methods require the preparation and measurement of samples of all or many possible combinations of candidates, which requires a lot of time, effort (e.g., the number of trials) and costs to prepare a combinatorial library. Furthermore, combinations other than the combinations of materials prepared as samples must be searched for using methods such as Bayesian optimization, which are commonly used in fields such as MI. However, in order to perform a sufficiently reliable search (prediction) using these conventional search methods, a large number of measurement results are required, making it difficult to sufficiently reduce costs such as time, effort (e.g., the number of trials) and costs.

[0023] Furthermore, as mentioned above, conventional methods using AI technology require a huge amount of training data to train the generative model, and the generative model must be learned over a long period of time, which means that obtaining an optimal solution requires a huge amount of time, effort, and expense.

[0024] Therefore, in the following embodiment, it is possible to create a combinatorial library that covers the entire search space from measurement data measured under a small number of experimental conditions, ranging from several to approximately 100, thereby making it possible to significantly reduce the cost required to obtain an optimal solution. Note that the optimal solution that the present invention searches for is not limited to the film formation conditions for the light-emitting layer, but may also be the film formation conditions for various thin films having certain characteristics.

[0025] Furthermore, the search space in this description may be a space representing n-dimensional experimental conditions to be searched, and may be a space in which a combinatorial library is created. The number of dimensions n of the search space may be the number of types of parameters adopted as experimental conditions. For example, if the adopted parameters are temperature and pressure, the number of dimensions may be two.

[0026] <1. System configuration example> Fig. 1 is a schematic diagram showing an example of the general configuration of an MI search system as an information processing system according to this embodiment. Fig. 2 is a block diagram showing an example of the general configuration of an experiment system according to this embodiment. Fig. 3 is a block diagram showing an example of the general configuration of a measurement system according to this embodiment. Fig. 4 is a block diagram showing an example of the general configuration of a prediction system according to this embodiment.

[0027] First, as shown in FIG. 1, the MI exploration system 1 is composed of an experiment system 10, a measurement system 20, and a prediction system 30.

[0028] Experimental System 10 As shown in FIG. 2, the experimental system 10 includes, for example, an operation terminal 11, a control system 12, and a film forming device 13, and produces a semiconductor wafer (hereinafter also referred to as a sample substrate) M having a light-emitting layer under specified experimental conditions.

[0029] Specifically, for example, the operation terminal 11 is configured with an information processing device having a communication function, such as a personal computer or a tablet terminal, and inputs experimental conditions S (hereinafter simply referred to as "next experimental conditions S") for creating the next sample substrate M to the control system 12. The operation terminal 11 may be communicably connected to the prediction system 30 via a predetermined network, such as the Internet or a LAN (Local Area Network). The experimental conditions S may be values ​​set by the user selecting or inputting various parameters into the operation terminal 11, or may be values ​​input from the prediction system 30.

[0030] The control system 12 is connected to the film formation apparatus 13, for example, via a wired or wireless line, and controls various parameters such as the film formation temperature, film formation time, composition, film thickness, chamber pressure, plasma excitation frequency, and dose amount set in the film formation apparatus 13 based on the input experimental conditions S.

[0031] The film formation device 13, for example, comprises a chamber equipped with a film formation process mechanism such as a turntable, and operates under control of the control system 12 to prepare a sample substrate M having a light-emitting layer with a layer structure to be measured. The prepared sample substrate M may have a light-emitting layer whose thickness and composition are uniform across the entire substrate, or may have a light-emitting layer whose thickness and composition vary depending on the position on the substrate, as prepared by conventional combinatorial techniques. For example, 16 rectangular regions arranged in a 4x4 pattern may be set on the film formation surface of a semiconductor wafer, and light-emitting layers with different thicknesses and compositions may be formed in each region. In this case, it is possible to prepare 16 samples at a time.

[0032] The sample substrate M produced in the experimental system 10 may be automatically or manually transported to the measurement system 20. If the film formation apparatus 13 has a function of measuring the sample substrate M in situ, the measurement system 20 in Fig. 1 may be omitted. In that case, measurement data R measured in the film formation apparatus 13 may be input to the prediction system 30 under the control of the control system 12.

[0033] Measurement System 20 As shown in FIG. 3, the measurement system 20 includes, for example, an operation terminal 21, a measurement device 22, and a storage device 23, and performs a predetermined measurement (for example, X-ray diffraction) on a sample substrate M set in the measurement device 22.

[0034] Specifically, for example, when the measurement technique is X-ray diffraction, the measurement device 22 may be an X-ray diffraction device.

[0035] The operation terminal 21 may be configured as, for example, an information processing device with a communication function, such as a personal computer or a tablet terminal. This operation terminal 21 is connected to, for example, the measuring device 22 via a wired or wireless line, and inputs measurement data R obtained by the measuring device 22 measuring the sample substrate M. Furthermore, the operation terminal 21 is communicably connected to the prediction system 30 via a predetermined network, such as the Internet or a LAN, and transmits the measurement data R input from the measuring device 22 to the prediction system 30.

[0036] The measurement data R sent from the operation terminal 21 to the prediction system 30 may be raw data obtained by the measurement device 22, or may be processed data (e.g., evaluation results) obtained by performing a predetermined processing on the raw data.

[0037] Furthermore, the operation terminal 21 may store the measurement data R and / or its processed data input from the measurement device 22 in the storage device 23. The storage device 23 may be, for example, an external storage device such as a hard disk drive, or may be a file server or cloud storage connected to the operation terminal 21 via a predetermined network.

[0038] Furthermore, the operation terminal 21 may be the same information processing device as the operation terminal 11 in the experimental system 10.

[0039] ·Prediction System 30 4, the prediction system 30 is composed of a cloud server 31, an AI server 32, and a cloud storage 33, and determines the experimental conditions S to be used in the next experiment (i.e., the fabrication of the next sample substrate M), and transmits the determined experimental conditions S to the experimental system 10. In this case, the prediction system 30 may determine the next experimental conditions S based on predicted values ​​(a prediction library, described later) obtained from measurement data R acquired in the previous experiment.

[0040] Specifically, the cloud server 31 is placed on the Internet (including a virtual private line (VPN) and the like), for example, and provides a cloud service for determining the experimental conditions S required for creating a combinatorial library. For example, when measurement data R is input from a service subscriber (user), the cloud server 31 inputs this to the AI ​​server 32. Then, the cloud server 31 transmits the experimental conditions S returned from the AI ​​server 32 in response to the input of the measurement data R to the operation terminal 11 of the experimental system 10 specified by the user.

[0041] The AI ​​server 32 is placed, for example, on the Internet (including a virtual private line (VPN) etc.), and when measurement data R is input from the API of the cloud server 31, it determines the next experimental conditions S and returns the determined experimental conditions S to the API.

[0042] When combinatorial methods are used to search for optimal solutions such as materials, compounds, or articles that are optimal for a certain application, in order to reduce the cost of the search and efficiently create a combinatorial library, it is desirable to set the next experimental condition S to an experimental condition for which measurements with similar parameters have not yet been performed.

[0043] Therefore, in this embodiment, the details of which will be described later, a method of determining the next experimental condition S may be used, for example, by setting a distance field for a search space corresponding to the combinatorial library to be created, and determining the next experimental condition S as an experimental condition located in an area of ​​the search space where measurements with similar parameters have not yet been performed (hereinafter also referred to as a sparse area).

[0044] Furthermore, when setting the distance field for the search space, the AI ​​server 32 calculates the characteristic distribution of the entire search space predicted from the measurement data R (for example, the distribution of X-ray diffraction data in this description, also referred to as a prediction library) as a bias (for example, the coefficient α(p k ) may also be used.

[0045] For example, the AI ​​server 32 may predict a characteristic distribution (prediction library) for the entire search space from a combination (hereinafter also referred to as a sample point) of one or more measurement data R input in the process of creating a combinatorial library and the experimental conditions S used to fabricate the sample substrate M from which each measurement data R was acquired, and use the predicted values ​​(hereinafter also referred to as characteristic predicted values) for each experimental condition in this prediction library as a bias when calculating the distance field to identify experimental conditions located in a sparse area where no experiments have yet been conducted nearby, and determine the identified experimental conditions as the next experimental conditions S.

[0046] In this way, by using the characteristic distribution (prediction library) of the entire search space predicted from the sample points in distance field calculation, it becomes possible to focus experiments on regions in the search space where an optimal solution is likely to exist (which may be regions where the optimal solution is highly likely to exist). In other words, it becomes possible to reduce the number of experiments in regions in the search space where an optimal solution is unlikely to exist (which may be regions where the optimal solution is unlikely to exist). This makes it possible to predict and create a highly accurate combinatorial library with fewer experiments, significantly reducing the time, effort (e.g., number of trials) and costs required to create a combinatorial library, and as a result, significantly reducing the cost required to obtain an optimal solution.

[0047] The prediction library can be generated using, for example, a generative model 34, the details of which will be described later. The generative model 34 may be configured to receive one or more experimental conditions as input and to output measurement data predicted for the experimental conditions as predicted characteristic values.

[0048] The determination of the next experimental condition S using the distance field may be performed by the cloud server 31 instead of the AI ​​server 32. In this case, the cloud server 31 may determine the next experimental condition S by using a prediction library generated by inputting the measurement data R to the AI ​​server 32 for distance field calculation.

[0049] Once the next experimental conditions S have been determined in the above manner, the cloud server 31 transmits the determined next experimental conditions S to the operation terminal 11 of the experimental system 10. As a result, an experiment based on the next experimental conditions S is carried out in the experimental system 10.

[0050] Furthermore, data such as the prediction library generated by the AI ​​server 32 (or cloud server 31) and the distance field of the entire search space when the previous experimental conditions S were determined may be stored in cloud storage 33 and used to determine the next experimental conditions S. Cloud storage 33 may be located, for example, on the Internet (including a virtual private line (VPN)). However, this is not limiting, and instead of cloud storage 33, a storage device connected to cloud server 31 or AI server 32 may be used.

[0051] 2. How to determine the experimental condition S using distance fields Next, a method for determining the next experimental condition S using the distance field will be described in detail below with reference to the drawings. In the following description, the experimental condition S will also be referred to as the experimental point S.

[0052] (If you do not use the prediction library) First, a method for determining the experimental point S when a prediction library is not used will be described. Fig. 5 is a flowchart showing an example of the experimental point determination operation when a prediction library according to this embodiment is not used. The operation illustrated in Fig. 5 may be executed, for example, by the AI ​​server 32 (or the cloud server 31) in the prediction system 30.

[0053] Step S101 As shown in FIG. 5, in step S101 of this operation, first, a plurality of candidate points p are arranged in a search space. In this description, the distribution of the plurality of candidate points p arranged in the search space is also referred to as a proposal distribution. Note that the candidate points p are points in an n-dimensional search space, as expressed by the following formula (1), and may be experimental conditions that may be selected as the experimental points S. In formula (1), pk denotes the position in the search space (i.e., the position of the candidate point), and R n denotes the n-dimensional search space.

number

[0054] Here, the distance field is a scalar field that represents the distance between the nearest experimental points among multiple experimental points, and can be expressed by the following equation (2). In equation (2), S j indicates experimental points that have already been tested.

number

[0055] Therefore, in this embodiment, as illustrated in Fig. 6, a plurality of candidate points p that can become the experiment point S are set in the search space, and the next experiment point S is determined from among these candidate points p. Fig. 6 is a diagram showing an example of a proposal distribution according to this embodiment.

[0056] Multiple candidate points p are arranged in the search space R n It is desirable to arrange the points so that the density is uniform or nearly uniform. A number of candidate points p that can become the experimental points S are arranged in the search space R n By arranging the candidate points p evenly within the range, it is possible to avoid selecting a candidate point p with parameters similar to an experimental point S that has already been tested as the next experimental point S, thereby further reducing the cost required to create a combinatorial library.

[0057] search space R n Candidate point p for k In the arrangement of the above, for example, as expressed by the following formula (3), one or more candidate points p j The candidate points p are arranged so that the minimum value of the distance field D(x) for each is maximized. k The position of r may be determined sequentially. i is the search space R nThe points may be randomly placed within the image, and the number of points may be large, for example 10,000.

number

[0058] According to these rules, the search space R n By placing the candidate point p within the search space R, as shown in Fig. 6, n A plurality of candidate points p can be arranged so that the density is uniform or approximately uniform with respect to the point p.

[0059] A search space R in which multiple candidate points p are placed n may be stored in the cloud storage 33, for example, and retrieved when necessary.

[0060] Step S102 Next, in step S102, as shown in FIG. 7, a search space R n An experiment point S0 at which an experiment is first performed (hereinafter also referred to as an initial experiment point) is determined from among a plurality of candidate points p arranged within Fig. 7. Fig. 7 is a diagram showing an example of candidate points selected in step S102 of Fig. 5.

[0061] The initial experiment point S0 may be selected at random from the candidate points p already placed, for example. However, this is not limiting, and for example, the initial experiment point S0 may be selected from the candidate points p placed first in creating the proposal distribution, or the candidate point p0 in the search space R n The candidate point p located closest to a predetermined position (e.g., the center) in k Various candidate points p may be selected, such as:

[0062] Step S103 In step S103, the initial experiment point S0 selected in step S102 is output as the next experiment condition S0. The output experiment condition S0 may be transmitted from the cloud server 31 to the experiment system 10. In response to this, the experiment system 10 fabricates a sample substrate M based on the received experiment condition S0. Then, the measurement system 20 measures the sample substrate M fabricated in the experiment system 10, thereby acquiring measurement data R0.

[0063] Step S104 In step S104, measurement data R0 obtained by measuring the sample substrate M based on the experimental conditions S0 is acquired from the measurement system 20. The measurement data R acquired by the measurement system 20 may be automatically transmitted from the operation terminal 21 to the cloud server 31, or may be manually transmitted to the cloud server 31 by the user from the operation terminal 21.

[0064] Step S105 In step S105, the search space R n For each candidate point p in , the distance field D(p) is calculated using equation (4) below:

number

[0065] By using equation (4), the value of the distance field D(p) becomes extremely small around an experimental point S where an experiment has already been conducted. If the value of equation (4) is large, it means that there are no experimental points S nearby that have already been tested, indicating that they have not yet been searched and therefore need to be searched. Conversely, if the value of equation (4) is small, it indicates that there are already searched points nearby and therefore the priority of measurement is low. Therefore, by determining the next experimental point S based on equation (4), it is possible to prevent an experimental point S close to an experimental point S that has already been tested, i.e., an experimental condition S with parameters similar to the experimental condition S that has already been tested, from being selected as the next experimental condition S. In other words, by selecting points in order from the largest D(p), it is possible to extract experimental conditions in order from locations that have not yet been searched.

[0066] Step S106 In step S106, by using the following equation (5), the candidate point p with the largest distance field D(p) calculated in step S105 is selected as the next experimental point S i is determined as follows.

number

[0067] Step S107 In step S107, similarly to step S103, the experimental point S selected in step S106 is i is the next experimental condition S i and may be transmitted to the experimental system 10. In contrast, in the experimental system 10, the experimental condition S i A sample substrate M is fabricated based on the above, and measurement is performed on the sample substrate M in the measurement system 20. As a result, measurement data R i is obtained.

[0068] Step S108 In step S108, similarly to step S104, the experimental condition S i Measurement data R obtained from sample substrate M based on i is obtained.

[0069] Step S109 In step S109, the search space R n It is determined whether sufficient searching has been performed for candidate points p. Whether sufficient searching has been performed may be determined, for example, based on whether a preset number of candidate points p have been selected, or whether the distance field D(p) at unselected candidate points p has become a sufficiently small value (for example, a value lower than a preset threshold value). Alternatively, the user may make a determination and input the result to the cloud server 31 from the operation terminal 21 of the measurement system 20. However, the determination in step S109 is not limited to these, and may be performed based on various methods.

[0070] As a result of step S109, the search space R n If it is determined that the search for is sufficient (YES in step S109), this operation ends. On the other hand, if it is determined that the search is not sufficient (NO in step S109), this operation returns to step S105, and the subsequent operations are executed.

[0071] 8 to 12 are diagrams showing examples of candidate points selected by repeatedly executing step S106 in Fig. 5. Fig. 8 shows an example in which up to the second candidate point S1 has been selected, Fig. 9 shows an example in which up to the third candidate point S2 has been selected, Fig. 10 shows an example in which up to the fourth candidate point S3 has been selected, and Fig. 11 shows an example in which up to the 16th candidate point S 16 Figure 12 shows an example of the case where up to the 100th candidate point S 99 An example is shown below where up to is selected.

[0072] As illustrated in FIGS. 8 to 12, in the process of repeating step S106 in FIG. 5, the candidate point p with the largest distance field D(p) is selected as the next experimental point S i By determining each time as n It is possible to place the experimental point S without any bias.

[0073] (When using the prediction library) Next, we will explain how to determine experimental points when using a prediction library. By using the prediction library obtained by the generative model 34, it becomes possible to prioritize points with large predicted characteristic values ​​as experimental targets. Therefore, when using a prediction library, a bias is applied to the predicted value in the distance field D(p) so that points with better predicted values ​​are more likely to be searched.

[0074] Fig. 13 is a diagram showing an example of a prediction library according to this embodiment. For clarity, in Fig. 13, Gaussians are used as model data instead of the prediction library obtained by the generative model 34.

[0075] In the example prediction library shown in Figure 13, the search space R n In the graph, there is a first distribution with a high peak value and a wide spread slightly above the center on the left side, and a second distribution with a higher peak value and a relatively narrow spread in the lower right corner. Distributions with these peaks are considered to be regions where an optimal solution is likely to exist (or regions where the optimal solution is likely to exist). The larger the predicted value, the more likely the optimal solution is to exist. This suggests that a more accurate combinatorial library can be generated by focusing experiments on regions with relatively large predicted values.

[0076] On the other hand, regions with small predicted values ​​are considered to be regions where an optimal solution is unlikely to exist (or may be regions where the optimal solution is unlikely to exist). Therefore, by reducing the number of experiments for regions with small predicted values, it is possible to significantly reduce the time and effort (e.g., number of trials) and costs required for creation of the prediction library while maintaining its accuracy.

[0077] Such a prediction library may be generated, for example, by inputting measurement data R into the generative model 34. The prediction library may be updated each time new measurement data R is acquired. Therefore, by conducting more experiments and accumulating more measurement data R, it is possible to generate a more accurate prediction library. Then, by using the updated prediction library in distance field calculations when determining the next experimental point, it is possible to determine a more accurate candidate point p as the next experimental point S.

[0078] If the predicted value is directly multiplied by the distance field D(p), excessive bias will be applied. Therefore, in this embodiment, a bias according to the digit of the predicted value is applied by taking the form of log(1+x), as shown in the following equation (6). Equation (6) is a conversion equation for converting the predicted value into a bias (coefficient α(p)) to be incorporated into the distance field calculation.

number

[0079] In equation (6), R(p) is a predicted value by the generative model 34. Furthermore, β may be a value calculated by the following equation (7). R(x) in equation (7) may be a value obtained by scoring the characteristic predicted values ​​under each experimental condition generated by the generative model 34.

number

[0080] The distance field (hereinafter referred to as the corrected distance field) D'(p k ) can be calculated using the following equation (8).

number

[0081] By calculating the corrected distance field D'(p) in this way, taking the predicted value into account, the distance field value will be larger for locations where the predicted value is estimated to be higher, making them more likely to be selected as the next experimental point S.

[0082] In equation (6), β is a constant, but its value may be set to the maximum value of R(p). If you want to prioritize search (search mode), increasing β will reduce the change in the log term, and the predicted value will not be prioritized as much. On the other hand, if you want to perform optimization-priority search (optimization mode), where you want only the optimal value rather than a broad search, setting β to a smaller value will enable a search that moves straight to the optimal solution.

[0083] 14 is a flowchart showing an example of an experiment point determination operation when using the prediction library according to this embodiment. The operation illustrated in FIG. 14 may be executed, for example, in the AI ​​server 32 (or the cloud server 31) in the prediction system 30, similar to the operation illustrated in FIG. 5. Furthermore, in the operation illustrated in FIG. 14, operations similar to those illustrated in FIG. 5 will be referred to and redundant explanations will be omitted.

[0084] Steps S101-S104 As shown in FIG. 14, in steps S101 to S104 of this operation, the search space R n Then, measurement data R0 for the initial experiment point S0 selected from the proposed distribution is acquired. The acquired measurement data R0 may be stored in, for example, cloud storage 33.

[0085] Step S201 Next, in step S201, the measurement data R acquired up to that point (measurement data R0 in the first case) is input to the generative model 34, thereby generating a prediction library.

[0086] Step S205 Next, in step S205, the search space R n For each candidate point p in, the modified distance field D'(p) is calculated using equations (6)-(8) above.

[0087] Step S206 In step S206, by using the following equation (9), the candidate point p with the largest corrected distance field D'(p) calculated in step S105 is selected as the next experimental point S i is determined as follows.

number

[0088] Steps S107-S109 Then, similar to steps S107-S109 in FIG. 5, the experimental point S i Measurement data for R i is obtained. Then, the search space R n It is determined whether a sufficient search has been performed for the above, and if it is determined that the search is sufficient, this operation is terminated, whereas if it is determined that the search is not sufficient, this operation returns to step S201 and the subsequent operations are executed.

[0089] In step S201 from the second time onwards, the measurement data R accumulated up to that point is used. 0- R i For example, the measurement data R0-R i may be fed back into the generative model 34 to generate an updated prediction library, or the measured data R i-1 The measurement data in the prediction library generated so far i The additional data may be input to the generative model 34 to generate an updated prediction library.

[0090] Fig. 15 is a diagram showing examples of candidate points selected in step S102 of Fig. 14. Figs. 16 to 20 are diagrams showing examples of candidate points selected in step S206 of Fig. 14, with Fig. 16 showing an example where up to the second candidate point S1 is selected, Fig. 17 showing an example where up to the third candidate point S2 is selected, Fig. 18 showing an example where up to the fourth candidate point S3 is selected, and Fig. 19 showing an example where up to the 16th candidate point S4 is selected. 16 Figure 20 shows an example of the case where up to the 100th candidate point S 99 An example is shown below where up to is selected.

[0091] First, as shown in FIG. 15, in step S102, the search space R n An experimental point S selected according to a predetermined rule from among a plurality of candidate points p arranged within may be determined as the initial experimental point S0.

[0092] Next, as illustrated in FIGS. 16 to 20, in the process of repeating step S206 in FIG. 14, the candidate point p with the largest corrected distance field D′(p) considering the prediction library is selected as the next experimental point S i By determining the value of each point, the search space R is expanded while selecting an experimental point S as far away as possible from the experimental point S that has already been tested. n Candidate points p in the area slightly above the center on the left side and in the area in the lower right corner, where the optimal solution is likely to exist, are actively determined as experimental points S. In other words, locations with larger corrected distance field D'(p) are sampled more frequently as the next experimental point S.

[0093] In this example, the location where the corrected distance field is large is selected as the next sampling candidate, and sampling (i.e., fabrication and measurement of the sample substrate M under the experimental conditions) is performed there, the predicted value by the generative model 34 is updated, and the next corrected distance field is recalculated. n This makes it possible to place the experimental points S in a region where the optimal solution is likely to exist (or in a region where the optimal solution is highly likely to exist) while suppressing bias, thereby enabling the process search to proceed quickly.

[0094] <3. Generative Model of Prediction Library> Next, the generative model (also called generative model) 34 according to this embodiment will be described in detail with reference to the drawings.

[0095] (When the decoder of the autoencoder is used as the generative model: First generative model) First, we will explain the case where a decoder of an autoencoder is used as a generative model.

[0096] An autoencoder with a general hourglass network structure has an input-side encoder and an output-side decoder, and after training, the decoder can be used alone as a generative model.

[0097] The input-side encoder has, for example, a convolutional neural network structure, inputs a high-dimensional observable variable (hereinafter simply referred to as a variable) x (hereinafter, the input-side variable x is also referred to as an input x), and outputs a low-dimensional latent variable z that is uniquely determined from this input x.

[0098] On the other hand, the decoder on the output side has, for example, a transposed convolutional neural network structure, and outputs a high-dimensional variable x reconstructed in response to an input of a low-dimensional latent variable z (hereinafter, the variable x on the output side is also referred to as the output x).

[0099] Autoencoders with such a structure are typically used in unsupervised learning, i.e., they can be trained without a supervised signal by training the output x to approach the input x.

[0100] In the learning of an autoencoder, it learns how to summarize to obtain a latent variable z that can recover the input x as the output x.

[0101] After the autoencoder has been trained, new data with the same dimensionality as the output x can be generated by inputting random values ​​with the same dimensionality as the latent variable z to the decoder. In other words, the encoder can be omitted and the decoder alone can function as a generative model.

[0102] Therefore, unlike conventional methods, we use the decoder in neural network generative models such as autoencoders as a standalone generative model, which makes it possible to obtain high-dimensional output from low-dimensional information while reducing the time and cost required to obtain an optimal solution.

[0103] When the decoder of the autoencoder is used as the generative model 34 (which will be referred to as the first generative model 100), the first generative model 100 may be a statistical model based on the joint probability distribution p(x|z) of the observable variable x and the latent variable (also referred to as the objective variable) z, and may be a neural network that can output new data by performing an operation using the latent variable on input data (for example, a transpose convolution operation, which will be described later). There are no restrictions on the latent variable, and various values ​​can be used.

[0104] Note that the neural network generative model in this description is not limited to the structure shown in the example, and various operations used in neural networks, such as CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), Transformer, Attention, etc. For clarity, the following description will be given using an example in which a decoder using a transposed convolution operation is used as the generative model.

[0105] Fig. 21 is a diagram showing a schematic configuration example of a first generative model according to this embodiment. As shown in Fig. 21, the first generative model 100 has a configuration consisting of a decoder (also called a generator), in other words, a configuration consisting of only a decoder with no encoder.

[0106] Then, by having the first generative model 100 learn the sample points, it becomes possible to generate X-ray diffraction data of a thin film (hereinafter also referred to as a measurement film) that can be formed under certain conditions without requiring an encoder. Here, the measurement film may be a thin film that is the target of characteristic prediction.

[0107] In this description, X-ray diffraction data (also called X-ray diffraction image) that can be obtained by XRD is used as an example of the predicted property, but the present invention is not limited to this and it is possible to configure the system to predict various properties. Furthermore, the experimental conditions in this description may be the experimental conditions used when forming the measurement film.

[0108] The input z is directly input to the decoder, which is the first generative model 100. Therefore, the input z to the first generative model 100 can be data of the same dimension as the latent variable z, which is lower in dimension than the output x.

[0109] Fig. 22 is a diagram for explaining the learning of the first generative model according to this embodiment. As shown in Fig. 22, the learning of the first generative model 100 uses a combination (i.e., sample points) of experimental conditions used when forming a certain measurement film and measurement data (here, X-ray diffraction data) obtained by measuring the measurement film formed under these experimental conditions.

[0110] In the independent learning of the first generative model 100, i.e., the decoder, experimental conditions S (input z) are input to the decoder. The learning unit (also called optimizer) 110 learns so that the output x from the first generative model 100 for this input z approaches the measurement data R actually obtained from a measurement film formed under the experimental conditions S.

[0111] In addition, various optimization algorithms 110 such as stochastic gradient descent (SGD), MomentumSGD, Nesterov's accelerated gradient algorithm (NAG), AdaGrad, RMSprop, AdaDelta, Adam, RMSpropGraves, SMORMS3, AdaMax, Nadam, Eve, Santa, GD by GD, AdaSecant, AMSGrad, AdaBound, and AMSBound can be used to train the first generative model 100 by the learning unit 110.

[0112] (When using Transformer as a generative model: second and third generative models) Next, we will explain some examples of using Transformer as a generative model 34.

[0113] Generative artificial intelligence (Generative AI) technology has made remarkable progress in recent years, and its range of applications is expanding. In particular, systems such as ChatGPT (registered trademark) have significantly improved their performance over the past few years, from basic sentence generation and summarization capabilities to advanced text generation capabilities that are indistinguishable from those created by humans. This has been achieved by using an extraordinary amount of training data and parameters.

[0114] Generative Pretrained Transformer (GPT) technology has the potential to apply its architecture to a variety of fields. The inventors have discovered that by applying this technology to the field of materials science (MI), it may be possible to build a system that can predict all physical properties and experimental values ​​in the future, even in areas where data is currently lacking.

[0115] The concept of this invention originated from a paper titled "Attention Is All You Need" published in 2017. This paper brought about a paradigm shift in AI technology, and since then, almost all AI technology has been based on the Transformer model.

[0116] The success of the Transformer model is based on the scaling law (see Non-Patent Document 3). Figure 23 is a graph showing the Transformer's scaling law. In Figure 23, the vertical axis represents performance (inversely), and the horizontal axis represents the amount of data. This property of infinitely improving performance as the amount of data increases has dramatically improved the capabilities of generative AI. To date, no upper limit has been found for performance as the amount of data and parameters increases.

[0117] Figure 24 is a diagram showing an example of the structure of a Transformer. The structure of a Transformer is a network as shown in Figure 24, and as the title "Attention Is All You Need" suggests, the core of the Transformer is an attention mechanism. This mechanism automatically focuses on important parts of the given input data and improves the quality of the output.

[0118] Predictive models in AI technology depend on how efficiently essential information can be extracted from given data. For example, in convolutional networks, crucial information is contained in the vicinity of pixels, so models can be constructed efficiently by extracting only the nearby information through filtering operations. On the other hand, in natural language processing, contextually related words can exist in physically distant locations, so how to capture the dependencies between distant locations is extremely important.

[0119] Such long-term dependencies in natural language have traditionally been handled using mechanisms with memory capabilities such as Long Short Term Memory (LSTM), but the invention of attention has made it possible to handle them more directly. The attention mechanism is expressed as follows:

number

[0120] Specifically, the attention mechanism evaluates the relationships between input data using three elements: query (Q), key (K), and value (V). This process calculates how each element of the input data is related to each other and generates a contextually meaningful response by selecting the most useful information for a particular query.

[0121] This mechanism enables generative AI to extract and generate accurate and relevant information even from text data with complex contexts and long-term dependencies. This technology is particularly important in the field of natural language processing, where it contributes to understanding deep contexts that are difficult to capture with conventional models and generating logical documents.

[0122] In order to incorporate these new technologies into the field of MI, the inventors developed the Combinatorial Generative Pretrained Transformer (CombiGPT), which combines the Transformer and combinatorial techniques.

[0123] In developing CombiGPT, the inventors used the C++ version of PyTorch to speed up the process and package it, and developed the Transformer part from scratch to make it easier to modify the structure and monitor variables.

[0124] Conventional combinatorial techniques aim to create many samples at once using masks and condition gradients and perform a comprehensive search. However, CombiGPT allows for virtually performing composition search mapping and condition search mapping similar to combinatorial searches by acquiring only a few data points. Furthermore, because predictions are performed using AI, it is possible to easily perform oxygen gas pressure gradient mapping and target-substrate distance gradient mapping, which were difficult to achieve with conventional combinatorial techniques. Therefore, in this embodiment, two types of CombiGPT will be described as examples.

[0125] Regular CombiGPT (second generative model) First, a case where a normal type CombiGPT is used as the second generative model will be described. Fig. 25 is a diagram showing a schematic configuration example of the second generative model according to this embodiment. Note that the normal type CombiGPT may be a generative model having a network structure similar to that of a general Transformer.

[0126] As shown in Figure 25, the second generative model 200 using the standard type of CombiGPT has an encoder-decoder model structure, with the left column in the drawing being the encoder 210 and the right column in the drawing being the decoder 220. When experimental conditions S are input from the encoder 210, the decoder 220 on the right side in the drawing generates the corresponding characteristic prediction value (in this example, X-ray diffraction data).

[0127] Specifically, the encoder 210 is composed of an input embedding layer 211, a position encoding layer 212, and a main part 213. Although not shown, a dropout layer may be added to each layer as appropriate.

[0128] The experimental conditions S input to the second generative model 200 are converted into vectors in the input embedding layer 211, and the input position (e.g., which parameter of the experimental conditions S each value represents) is added in the position coding layer 212, after which the input is input to the main part 213.

[0129] The main part 213 has, for example, a hierarchical structure of N layers (N is a natural number), and processing is repeated in each layer. Each layer of the main part 213 has two sub-layers: a multi-head attention layer 214 and a position-wise feed-forward network layer 216, and each sub-layer is followed by add & norm layers 215 and 217 that perform residual connection and normalization processing. Data that has undergone processing for N layers of the main part 213 is input to a decoder.

[0130] The decoder 220 is composed of an input embedding layer 221, a positional coding layer 222, a main part 223, a fully connected (Linear) layer 230, and a softmax layer 231. The processing from the input embedding layer 221 to the positional coding layer 222 may be the same as the processing from the input embedding layer 211 to the positional coding layer 212 in the encoder 210. Furthermore, the number of layers in the main part 223 may be the same as the number of layers in the main part 213 of the encoder.

[0131] Each layer of the main part 223 has two sub-layers (a multi-head attention layer 224 and a position-wise forward propagation network layer 228, and two summation and normalization layers 225 and 229) similar to those of the main part 213 of the encoder 210, as well as a multi-head attention layer 226 that receives the output from the first sub-layer, the multi-head attention layer 224, and a summation and normalization layer 227 that is placed after the multi-head attention layer 226. The first multi-head attention layer 224 is responsible for masking information from the position coding layer 222.

[0132] In the above configuration, the encoder 210 and the decoder 220 can be replaced with specialized encoders. This allows for flexible configuration changes depending on the type of experiment to be predicted. Since X-ray diffraction patterns are typically acquired in typical thin film experiments, this description will exemplify a decoder 220 that generates X-ray diffraction data PV as characteristic prediction values, taking into account factors such as ease of data collection, and an encoder 210 that receives experimental conditions including temperature and oxygen pressure as input.

[0133] The second generative model 200 using the normal type of CombiGPT can be constructed by transfer learning a pretrained transformer 900 that has been pretrained with a huge amount of training data 990, as illustrated in Fig. 26. The network structure of the pretrained transformer 900 may be similar to that of the second generative model 200 shown in Fig. 25.

[0134] Lightweight CombiGPT (the third generative model) Next, a case where a lightweight CombiGPT is used as a third generative model will be described. FIG. 27 shows a schematic configuration example of the third generative model according to this embodiment. In the following description, components similar to those of the second generative model 200 illustrated in FIG. 25 are denoted by the same reference numerals, and redundant description will be omitted. Note that the lightweight CombiGPT may be a generative model having a network structure that simplifies the network structure of a general Transformer to reduce processing load.

[0135] As shown in Figure 27, the third generative model 300, which uses the lightweight CombiGPT, does not use an encoder-decoder system, and generates X-ray diffraction data PV, which is the characteristic prediction value, directly from the vector of experimental conditions S without passing through latent variables. Since this corresponds to a pattern that uses only an encoder, its structure is similar to GPT1. However, its usage is similar to BERT (Bidirectional Encoder Representations from Transformers), which also uses an encoder type.

[0136] In other words, this type of encoder-type generative model defines the latent variable vector as the output data itself, making its mechanism very simple and enabling highly accurate analysis even with lightweight and small amounts of data.

[0137] 28 is a graph for explaining the learning status of the third generative model according to this embodiment. As shown in FIG. 28, the learning of the third generative model 300 using the lightweight CombiGPT progressed smoothly with increasing number of times, and after about 12,000 times or more, the learning progressed to 10 -3 Learning progressed to level loss.

[0138] CombiGPT (the fourth generative model) using GPT-2 While the second generative model 200 and the third generative model 300 described above are based on GPT-1, this embodiment is not limited to this and may be based on an advanced version of GPT such as GPT-2, GPT-3, GPT-3.5, or GPT-4. Here, we will explain the case where a standard type CombiGPT based on GPT-2 is used as the fourth generative model 400.

[0139] Fig. 29 is a diagram showing a schematic configuration example of a fourth generative model according to this embodiment. In Fig. 29, the same components as those in the second generative model 200 shown in Fig. 25 are denoted by the same reference numerals, and redundant explanations will be omitted.

[0140] As shown in Figure 29, the fourth generative model 400 has the structure of an encoder-decoder model, similar to the second generative model 200 illustrated in Figure 25, with the left column in the drawing being the encoder 410 and the right column in the drawing being the decoder 420.

[0141] However, in the fourth generative model 400, in the configurations of the encoder 410 and the decoder 420, the summation and normalization layers 215, 225, and 227 arranged at the outputs of the multi-head attention layers 214, 224, and 226 are replaced with normalization (Norm) layers 411, 421, and 423 arranged at the input of the multi-head attention layer 214 and addition (Add) layers 412, 422, and 424 arranged at the output. Similarly, the summation and normalization layers 217 and 229 arranged at the outputs of the position-by-position forward propagation network layers 216 and 228 are replaced with normalization (Norm) layers 413 and 425 arranged at the inputs of the position-by-position forward propagation network layers 216 and 228 and addition (Add) layers 414 and 426 arranged at the outputs. Furthermore, in the fourth generative model 400, a normalization layer 430 is added to the output of the decoder 420, i.e., the input of the fully connected (Linear) layer 230.

[0142] As such, CombiGPT used as the generative model 34 of this embodiment is not limited to GPT-1, and can be configured based on an advanced version of GPT such as GPT-2, GPT-3, GPT-3.5, or GPT-4. Furthermore, even when based on an advanced version of GPT such as GPT-2, GPT-3, GPT-3.5, or GPT-4, it is not limited to a standard type that uses an encoder-decoder system, such as the second generative model 200 or the fourth generative model 400, and can also be a lightweight type that uses the encoder side, such as the third generative model 300.

[0143] - Prediction results from generative models Fig. 30 is a graph showing the prediction results of the third generative model after training according to this embodiment. As shown in Fig. 30, the third generative model 300 after training is able to reproduce almost all data. In Fig. 30, light-colored areas indicate characteristic prediction values ​​by the third generative model 300, and dark-colored areas indicate areas in the actual measurement data that deviate from the characteristic prediction values.

[0144] Next, cross-validation was performed to evaluate general performance. Figure 31 is a diagram explaining the concept of this cross-validation method. As shown in Figure 31, this method involves excluding one item from the acquired dataset and isolating the item so that it cannot be recognized by the AI ​​system. After that, learning is carried out using the remaining data, and predictions are then made for the excluded data item, and the degree of agreement between the prediction results and the actual data is verified. This method is recognized as being particularly effective when the size of the dataset is limited.

[0145] In the performance evaluation of the third generative model 300 using the cross-validation method, one of the 16 sample points is removed, and the remaining 15 sample points are used to train the third generative model 300. Then, the performance of the third generative model 300 is evaluated by repeatedly comparing the measurement data of the removed sample point with the predicted characteristic values ​​obtained by inputting the experimental conditions of this sample point into the third generative model 300, while changing the sample point to be removed.

[0146] Fig. 32 is a diagram illustrating the results of evaluating the third generative model according to this embodiment using the cross-validation method. As shown in Fig. 32, even in the performance evaluation using the cross-validation method, it was found that although there were some areas where quantitative reproduction of peak intensity was poor, the position and presence or absence of peaks could be reproduced almost entirely. This level is sufficient for practical use, and can be said to be quite accurate.

[0147] 33 to 36 are diagrams showing examples of a prediction library obtained by performing combinatorial mapping of a location where no experiment was performed using the third generative model according to this embodiment. -4 Figure 34 shows the temperature combination when the pressure is 2.64 × 10 Torr. -3 FIG. 35 shows the temperature combination when the temperature is 115.2°C, and FIG. 36 shows the pressure combination when the temperature is 224.4°C. FIG. 37 shows an example of a prediction library obtained by continuously changing the temperature and pressure input to the third generative model according to this embodiment. In FIG. 37, (a) shows a predicted X-ray analysis image for the entire measurement angle θ, (b) shows a two-dimensional graph of the temperature combination and pressure combination at angle θ = 30.74°C, and (c) shows a two-dimensional graph of the temperature combination and pressure combination at angle θ = 37.46°C.

[0148] Creating an actual combinatorial sample with a temperature gradient requires preparations such as temperature calibration, which is extremely costly. In contrast, as shown in Figures 33 to 37, when using the third generative model 300 according to this embodiment, it is possible to easily predict and create an entire combinatorial library simply by inputting data, which is thought to significantly accelerate process optimization. Furthermore, since it is impossible to actually create a combinatorial sample with a pressure gradient, this optimization can only be achieved with CombiGPT, and it can be seen that there are no restrictions on the experimental conditions that can be achieved. This is true not only for the third generative model 300, but also for the first generative model 100 and the second generative model 200.

[0149] <4. Generation of prediction library> Next, generation of a prediction library using the generative model 34 according to this embodiment will be described.

[0150] As described above, the trained generative model 34 can predict measurement data (e.g., X-ray diffraction data) of a measurement film formed under certain experimental conditions by inputting those experimental conditions. Therefore, according to this embodiment, by inputting the experimental conditions of the entire search space into the first generative model 100, it is possible to interpolate measurement data under experimental conditions for which no actual measurement data has been obtained, and create a prediction library for the entire search space.

[0151] 38 is a flowchart showing an example of the operation of the generative model according to this embodiment when creating a prediction library for the entire search space. As shown in FIG. 38, in this operation, first, one experimental condition in the search space is input to the trained generative model 34 (step S1001). In response, the generative model 34 outputs a predicted characteristic value predicted for the measurement film formed under the input experimental condition (step S1002).

[0152] Next, it is determined whether all experimental conditions within the search space have been input into the generative model 34 (step S1003), and if there are experimental conditions that have not yet been input (NO in step S1003), the operation returns to step S1001 and the subsequent operations are repeated.

[0153] In this way, when all experimental conditions in the search space are input to the generative model 34 and the predicted characteristic values ​​for each experimental condition are obtained (YES in step S1003), the predicted characteristic values ​​for all experimental conditions are collected to create a prediction library for the entire search space (step S1004). After that, this operation ends.

[0154] <5. Hysteresis due to other film formation processes> FIG. 39 is a diagram showing an example of a sample substrate according to this embodiment. In the combinatorial method, in consideration of search efficiency, for example, multiple rectangular regions are set on the film formation surface of a single semiconductor wafer, and measurement films with different film thicknesses, compositions, etc. are formed in each region. For example, as illustrated in FIG. 39, 16 rectangular regions #1 to #16 arranged in a 4×4 pattern may be set on the film formation surface of the semiconductor wafer W, and measurement films with different film thicknesses, compositions, etc. may be formed in each region. In this case, it is possible to produce 16 samples at a time.

[0155] However, in the film formation process, it is not possible to simultaneously form thin films in two or more regions (two or more regions in rectangular regions #1 to #16) with different experimental conditions. Therefore, the measurement film formed first may be damaged by process changes such as temperature and pressure changes when forming measurement films in other regions later, and the characteristics may change from the characteristics at the time of film formation (hysteresis due to other film formation processes). For example, the characteristic q1 of the measurement film formed first may be significantly different from the characteristics at the time of film formation due to process damage caused by the formation of the other 15 measurement films. Equation (11) expresses the characteristic q taking into account the hysteresis of each of the 16 samples formed on one semiconductor wafer W.

number

[0156] To address such problems, it is possible to use the Transformer's position coding layer (e.g., the position coding layer 212 in the second generative model 200 and the third generative model 300 of this embodiment) to treat the deposition order of each measurement film as a sequence dimension.

[0157] That is, the position coding layer of the Transformer adds position information to data arranged in the sequence dimension, as shown in the following equations (12) and (13).

number

number

[0158] Therefore, as illustrated in Figure 40, for example, by arranging sample points (combinations of experimental conditions and measurement data) of each measurement film in the sequence direction according to the film formation order and inputting them into the generative model 34 (the second generative model 200 or the third generative model 300), position information (position encoding) according to the film formation order is added to each sample point in the generative model 34, making it possible to generate a predictive library that takes into account hysteresis that occurs depending on the film formation order.

[0159] On the other hand, if hysteresis is not taken into consideration, for example, as shown in Fig. 41, the sample points of each measurement film can be arranged in the batch direction and input to the generation model 34. In this case, the same position information (e.g., position encoding = 0) is added to each sample point, making it possible to generate a prediction library that does not take hysteresis into consideration. Note that Fig. 40 is a diagram showing an example of an input data arrangement when hysteresis is taken into consideration according to this embodiment, and Fig. 41 is a diagram showing an example of an input data arrangement when hysteresis is not taken into consideration according to this embodiment.

[0160] <6. Examples of combinatorial libraries> Next, some examples of combinatorial libraries that can be generated by the generative model 34 according to this embodiment will be described.

[0161] Fig. 42 is a diagram showing an example of a combinatorial library obtained by a generative model according to this embodiment. For the sake of simplicity, (b) to (e) of Fig. 42 show combinatorial libraries at a certain angle (θ2), but in reality, a three-dimensional combinatorial library with the angle (θ) as the z-axis may be generated.

[0162] Figure 42(a) shows an example of measurement points input to the generative model 34, and Figure 42(b) to (e) show examples of combinatorial libraries at a certain angle (θ2) obtained for the input of Figure 42(a). In Figure 42, (b) is a graph of the temperature gradient versus the composition gradient, (c) is a graph of the composition gradient versus the composition gradient, (d) is a graph of the pressure gradient versus the target-substrate distance, and (e) is a graph of the pressure gradient versus the composition gradient.

[0163] As shown in (a) and (b) of FIG. 42, the generation model 34 according to this embodiment can generate a combinatorial library of temperature gradients versus composition gradients (see (b)) by acquiring X-ray diffraction data at a small number of experimental points in the search space (see (a)). Furthermore, as shown in (a) and (c) of FIG. 42, a combinatorial library of dual composition gradients (composition gradients versus composition gradients), which is a strength of conventional combinatorial methods, can also be estimated from a small number of sample points. Furthermore, as shown in (a), (d), and (e) of FIG. 42, it is also possible to generate a peak intensity map versus pressure gradient, which was difficult with conventional combinatorial methods. For example, it is possible to generate a combinatorial library in which the vertical axis represents the pressure gradient and the horizontal axis represents the target-substrate distance gradient, as shown in (d), or a combinatorial library in which the vertical axis represents the pressure gradient and the horizontal axis represents the composition gradient, as shown in (e). In addition, it is possible to create combinatorial libraries with the vertical axis representing the pressure gradient and the horizontal axis representing the temperature gradient, which was not possible with conventional combinatorial methods, and to create peak intensity mappings in which the target-substrate distance gradient and the pressure gradient are changed simultaneously.

[0164] As described above, by using the generation model 34 according to this embodiment, it is possible to establish a technique for mapping peak intensities in X-ray diffraction data obtained by XRD.

[0165] <7. Summary> As described above, according to this embodiment, by determining the next experimental conditions S based on the distance field, it is possible to avoid selecting experimental conditions with parameters similar to experimental conditions S that have already been tested as the next experimental conditions S, thereby making it possible to significantly reduce the time, effort (e.g., number of trials) and costs required to create a combinatorial library.

[0166] Furthermore, by using the characteristic distribution (prediction library) predicted by the generative model 34 as a bias when calculating the distance field, it is possible to focus experiments on areas in the search space where an optimal solution is likely to exist (which may be areas where the optimal solution is highly likely to exist), while reducing the number of experiments in areas where the optimal solution is unlikely to exist (which may be areas where the optimal solution is unlikely to exist), thereby further reducing the cost required to create a combinatorial library.

[0167] The next experimental conditions S can be determined using, for example, an Acquisition Function in the context of Bayesian optimization, but in this embodiment, an original algorithm that is easier to use is proposed. As described above, this algorithm is configured to determine the next experimental conditions S by preferentially searching for locations that are likely to have good predicted values ​​by the generative model 34, while minimizing overlap with experimental conditions S that have already been investigated. Specifically, a field called a distance field is introduced to distinguish between experimental conditions S that have already been investigated and new experimental conditions S. This field exists in a space whose dimensions are the number of parameters that make up the experimental conditions to be determined, and is defined by the above-described formulas (1) to (4).

[0168] Furthermore, according to the MI search system 1 of this embodiment, it is possible to automatically run a loop (also called a search loop) from determining the experimental conditions S to producing and measuring the sample substrate M, updating the predicted values, and determining the next experimental conditions S taking the predicted values ​​into consideration.

[0169] For example, an automatic thin film deposition device (jointly developed by the present inventors and Vacuum Products Co., Ltd.) that automatically creates the next sample substrate M based on input experimental conditions S is introduced as the film deposition device 13 of the experimental system 10, Microsoft (registered trademark) Azure is introduced as the cloud server 31 in the prediction system 30, and a deep learning server from SCT Corporation is introduced as the AI ​​server 32. Then, by setting the next experimental conditions S to be automatically downloaded in response to a request from the film deposition device 13, the operation flow shown in Fig. 14 is configured to be automatically executed.

[0170] As described above, this embodiment makes it possible to realize an MI exploration system 1 that incorporates the latest AI technology to explore materials and optimize processes. While large amounts of big data are required to run AI, simplifying the structure and mechanism of the generative model 34 makes it possible to operate with less data and achieve sufficient accuracy. Furthermore, because data uploaded to the prediction system 30 is automatically accumulated, it is also possible to use a pre-trained GPT as the generative model 34.

[0171] <8. Hardware Configuration> The operation terminals 11 and 21, control system 12, measurement device 22, cloud server 31, AI server 32, and generative model 34 according to the above-described embodiments and their modifications can be realized by an information processing device 1000 having a configuration such as that shown in FIG. 43. FIG. 43 is a hardware configuration diagram showing an example of the information processing device 1000 that realizes these functions. The information processing device 1000 includes a CPU 1100, a ROM (Read Only Memory) 1200, a RAM (Random Access Memory) 1300, a recording device 1400, an input / output interface (I / F) 1500, and a communication unit 1600. The various units of the information processing device 1000 are connected by a bus 1700.

[0172] The CPU 1100 operates and controls each unit based on programs stored in the ROM 1200 or the recording device 1400. For example, the CPU 1100 loads the programs stored in the ROM 1200 or the recording device 1400 into the RAM 1300 and executes processing corresponding to the various programs.

[0173] The ROM 1200 stores boot programs such as a BIOS (Basic Input Output System) that is executed by the CPU 1100 when the information processing device 1000 is started up, programs that depend on the hardware of the information processing device 1000, and the like.

[0174] Recording device 1400 is a computer-readable recording medium that non-temporarily records programs executed by CPU 1100 and data used by such programs, and is a storage device that uses an HDD (Hard Disk Drive) or a CD (Compact Disk). Specifically, recording device 1400 is a recording medium that records programs for executing the operations of the present invention, which are an example of program data.

[0175] The communication unit 1600 is an interface for connecting the information processing device 1000 to an external network 1650 (e.g., the Internet). For example, the CPU 1100 receives data from other devices and transmits data generated by the CPU 1100 to other devices via the communication unit 1600.

[0176] The input / output I / F 1500 is an interface for connecting the input / output device 1650 and the information processing device 1000. For example, the CPU 1100 receives data from an input device such as a keyboard or a mouse via the input / output I / F 1500. The CPU 1100 also transmits data to an output device such as a display, a speaker, or a printer via the input / output I / F 1500. The input / output I / F 1500 may also function as a media interface for reading programs and the like recorded on a predetermined recording medium.

[0177] For example, when the information processing device 1000 realizes the functions of the operation terminals 11 and 21, the control system 12, the measurement device 22, the cloud server 31, the AI ​​server 32, and the generative model 34 according to the above-described embodiments, the CPU 1100 of the information processing device 1000 realizes these functions by executing programs loaded onto the RAM 1300. The recording device 1400 stores the programs according to the present invention. The CPU 1100 reads and executes program data from the recording device 1400, but as another example, the CPU 1100 may obtain these programs from other devices via an external network 1650.

[0178] Although the embodiments and modifications of the present invention have been described above, the technical scope of the present invention is not limited to the above-described embodiments or modifications thereof, and various modifications are possible within the scope of the gist of the present invention. Furthermore, components from different embodiments and modifications may be combined as appropriate.

[0179] Furthermore, the effects of the embodiments and their modifications described in this specification are merely examples and are not intended to be limiting, and other effects may be achieved. [Explanation of symbols]

[0180] 1 MI Search System 10 Experimental System 11, 21 Operation terminal 12 Control System 13 Film deposition equipment 20 Measurement System 22 Measuring equipment 23 Storage device 30 Prediction System 31 Cloud Server 32 AI Server 33. Cloud Storage 34 Generative Model 100 First Generative Model 110 Learning Department 200 Second Generative Model 210 Encoder 211, 221 Input Embedding Layer 212, 222 Position Encoding layer 213, 223 Main part 214, 224, 226 Multi-Head Attention Layer 215, 217, 225, 227, 229 Add & Norm Layer 216, 228 Position-wise Feed-Forward Network Layer 220 decoder 230 Fully connected (Linear) layer 231 Softmax Layer 300 The Third Generative Model M sample board R measurement data S Experimental conditions

Claims

1. An information processing method for creating a combinatorial library of a search space whose dimensionality is the number of parameters constituting experimental conditions, comprising: generating a prediction of a distribution of a property in the search space; calculating a distance field from one or more experimental conditions in the search space that have been tested, taking into account the predicted values; determining the following experimental conditions based on the calculated distance field; Obtaining measurement results obtained by measuring a sample prepared using the determined next experimental conditions; The predicted value is updated based on the acquired measurement results. An information processing method in which the above is repeated one or more times.

2. setting a plurality of candidate points arranged uniformly or approximately uniformly in the search space; the distance field is calculated for each of the plurality of candidate points; The next experimental condition is that the candidate point with the largest value of the distance field is determined from among the plurality of candidate points. The information processing method according to claim 1 .

3. Let p be the position of each candidate point in the search space, α be a coefficient calculated from the predicted value at each candidate point, and S be the position of an experimental condition that has been tested in the search space. The distance field D′(p) is calculated using the following equation (1): [Equation 1] The information processing method according to claim 2 .

4. Let R be the value obtained by scoring the predicted value and β be a constant, then the coefficient α is calculated using the following equation (2): [Equation 2] The information processing method according to claim 3 .

5. The information processing method according to claim 4 , wherein the constant β is a maximum value of the value R.

6. 2. The information processing method according to claim 1, wherein the predicted value is generated using a generative model that takes as input a combination of experimental conditions and measurement results obtained by measuring a sample prepared using the experimental conditions, and outputs the predicted value.

7. The information processing method according to claim 6 , wherein the generative model is a Transformer.

8. The generative model is An input embedding layer located at the input stage; a position encoding layer that adds position information to the values ​​output from the input embedding layer; a main part arranged at the output of the position coding layer and having a hierarchical structure of N layers (N is a natural number); A fully connected (Linear) layer is placed at the output of the main part; a softmax layer disposed at the output of the fully connected layer; It consists of Each layer of the main part includes two sub-layers, a multi-head attention layer and a position-wise feed-forward network layer, and an add & norm layer that performs residual connection and normalization processing after each of the sub-layers. The information processing method according to claim 6.

9. The information processing method according to claim 6 , wherein the generative model is a decoder portion of an autoencoder.

10. The sample includes a plurality of regions prepared under a plurality of different experimental conditions; the measurement is performed for each of the plurality of regions; The measurement results of each of the plurality of regions are arranged in a sequence dimension and input to the generative model.

9. The information processing method according to claim 7 or 8.

11. The information processing method according to claim 1 , further comprising: preparing the sample using the determined next experimental conditions.

12. 2. The information processing method according to claim 1, wherein the sample is a light-emitting layer made of one or more thin films.

13. An information processing device connected to an experimental system for preparing a sample to be measured and a measurement system for measuring the sample via a predetermined network, generating a predicted value of the characteristic distribution in a search space whose dimensionality is the number of parameters constituting the experimental conditions for producing the sample; calculating a distance field from one or more experimental conditions in the search space that have been tested, taking into account the predicted values; determining the following experimental conditions based on the calculated distance field; transmitting the determined next experiment conditions to the experiment system; a measurement result obtained by measuring a sample prepared in the experimental system using the next experimental condition in the measurement system; The predicted value is updated based on the acquired measurement results. and repeating this one or more times to create a combinatorial library of the search space.

14. A program for causing a processor included in an information processing device connected to an experimental system for preparing a measurement target sample and a measurement system for measuring the sample via a predetermined network, the program comprising: generating a predicted value of a characteristic distribution in a search space whose dimensionality is the number of parameters constituting the experimental conditions for producing the sample; calculating a distance field from one or more experimental conditions in the search space that have been tested, taking into account the predicted values; determining next experimental conditions based on the calculated distance field; transmitting the determined next experiment conditions to the experiment system; acquiring measurement results obtained by measuring, in the measurement system, a sample prepared in the experimental system using the next experimental conditions; updating the predicted value based on the acquired measurement results; A program for causing the processor to execute the above.

Images